aihomelabprivacy

Run MiniMax H3 in ComfyUI: Local Audio & 2K Video

🤖 Researched and drafted automatically from the official docs, and reviewed before publishing. Commands are taken from the source projects — but always sanity-check before running anything on your own hardware.

MiniMax H3 is a unified model that generates audio and video together in a single pass—no separate synthesis steps, no API fees, all on your own hardware. ComfyUI’s modular node graph makes it straightforward to build and iterate on audio-video workflows without touching code. If you’ve got a decent GPU and want to produce synchronized media locally, this is the fastest path.

What You’ll Need

MiniMax H3 runs on mid-range consumer GPUs. Minimum specs: 8GB VRAM, 16GB system RAM. Optimal: 12GB+ VRAM for 2K generation without long waits. You’ll also need enough disk space for model weights (roughly 20–30GB for the full checkpoint) plus scratch space for output frames.

Step 1: Install ComfyUI

Start with a clean Python 3.11+ environment on Windows, Linux, or macOS. Clone the main repository:

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

Run the setup script for your OS. On Linux or macOS:

./install.sh

On Windows, use the provided batch file or follow the manual install steps in the repository. Once installed, launch ComfyUI:

python main.py

The UI will open at http://127.0.0.1:8188 by default. Keep this on your local network or behind a VPN—never expose ComfyUI to the public internet.

Step 2: Download the MiniMax H3 Model

The model weights are hosted on Hugging Face. You’ll need to accept the license on the model card, then download the checkpoint. Place it in your models directory:

mkdir -p models/checkpoints
cd models/checkpoints
# Download MiniMax H3 checkpoint (use huggingface-cli or manual download)
# The exact filename and size depend on the released version
cd ../..

ComfyUI will auto-detect models in the standard models/ folder structure. If you want to organize differently, edit extra_model_paths.yaml to point to custom locations.

Step 3: Load or Create a MiniMax H3 Workflow

Inside ComfyUI, you have two options:

Option A: Use a pre-built template

Visit https://comfy.org/workflows/ and search for “MiniMax H3” or “audio video”. Download a template workflow (.json file) and load it into ComfyUI via Ctrl + O.

Option B: Build from scratch

In the ComfyUI canvas, right-click to add nodes. You’ll need:

  1. A text prompt node (for the video description)
  2. A text prompt node for audio description (if separate from video)
  3. The MiniMax H3 node (once you’ve installed or created the custom node)
  4. A VAE decoder to convert latents to video frames
  5. An audio output node
  6. A video output node

The exact node names and connections depend on whether MiniMax H3 is packaged as a built-in node or a custom node. Check the ComfyUI docs or community repositories for the latest MiniMax integration.

Step 4: Configure Generation Parameters

In the MiniMax H3 node, you’ll set:

  • Resolution: typically 1280×720 for 2K (or 1920×1080 for full HD)
  • Duration: video length in seconds (default often 6–10 seconds)
  • Seed: for reproducibility
  • Guidance scale: controls how strongly the model adheres to your prompt
  • Audio sample rate: usually 16kHz or 24kHz

Start with conservative settings (6–8 second clips, guidance 7.5) to get a feel for generation time on your hardware.

Step 5: Generate and Queue

Fill in your text prompts. For best results, describe both the visual scene and the audio mood or sounds you want. Example:

  • Video prompt: “A serene forest with sunlight filtering through leaves, soft wind rustling branches”
  • Audio prompt: “Ambient forest soundscape with gentle wind and bird chirps”

Press Ctrl + Enter to queue the workflow. ComfyUI will load the model into VRAM, run inference, and save outputs to output/ by default. Depending on your GPU and settings, a 6-second 2K clip takes 2–10 minutes.

Step 6: Review and Iterate

Once generation finishes, preview the video and audio in the output folder. Tweak prompts, guidance, or seed and re-queue. ComfyUI’s asynchronous queue lets you stack multiple jobs and walk away.

For production workflows, use App Mode to expose your workflow through a simpler UI, or integrate the API endpoints into your own application.

Keeping It Private

ComfyUI runs entirely offline once models are cached. Never port-forward the default 8188 port. If you need remote access, use a VPN tunnel or SSH port forwarding. The server binds to 127.0.0.1 by default; if you need LAN access, bind to your machine’s local IP explicitly with --listen 192.168.x.x (adjust to your subnet).

Is It Worth It?

Yes, if you own the hardware and want to iterate fast without per-minute API charges. MiniMax H3 produces synchronized audio and video in one shot—no latency between synthesis steps. The trade-off is upfront GPU investment and some patience during generation. For content creators, researchers, or anyone building automated workflows, the control and cost savings justify the setup. For occasional one-off videos, cloud APIs might be simpler. But on your own box, you own the output, the data, and the workflow forever.

Related guides

← All guides